We've been testing vision-language models for parsing complex PDFs (tables, charts, handwriting). Results are promising but inconsistent — great on clean docs, shaky on messy scans. Is anyone running this in production at volume? What accuracy are you seeing, and how do you handle the long tail?